IQuest-Q1 320B MoE model gets day-0 vLLM support
The engineering detail is that vLLM needed no new core code — the model runs on a hybrid KV cache layout where only 25 of 88 layers scale with context, cutting memory demands for very long prompts. Teams self-hosting coding agents can serve a 320B-class model on fewer GPUs than dense equivalents require. Compared with dense models of similar capability, the 15B active-parameter design trades throughput costs for much lower per-token compute.
IQuest-Q1 is a 320B-parameter MoE model from IQuest Research with 15B active parameters per token, built for agentic coding.
It uses 256 experts with 8 live per token and supports a 524,288-token context window.vLLM shipped day-0 support by combining existing components: a hybrid KV cache coordinator that applies a 3-sliding-to-1-full attention layer mix so only 25 of the 88 layers grow cache with context, a sinks path in the attention backends for the checkpoint's sink attention, and EAGLE speculative decoding with probabilistic draft sampling via the model's recursive MTP head.
IQuest-Q1 by @IQuest_research has day-0 support in vLLM: a 320B MoE for agentic coding, 15B active per token, 256 experts with 8 live, 524,288 context. 🎉
Built on pieces vLLM already has: 🔹 The hybrid KV cache coordinator, for a 3 sliding to 1 full attention layer mix, so only 25 of the 88 layers grow a cache with the context 🔹 A sinks path in the attention backends, for the sink attention the checkpoint asks for 🔹 EAGLE speculative decoding with probabilistic draft sampling, which is how the recursive MTP head drafts
This report is categorized under AITOP's "Models" desk, focusing on China's AI industry and domestic foundation model ecosystem. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows and US-China foundation model benchmarks and commercialization comparison. Key entities and technical topics under continuous tracking include IQuest-Q1, vLLM, open-source models, agentic coding.
Source attribution: curated from vLLM (@vllm_project).
